Embedded Space Based Dimensionality Reduction Modeling Research And Its Applications

In this paper, on the bases of machine learning theory, by introducing the ideas of the locality, pairwise constraints, semi-supervised learning and kernel mapping, we discuss the feature selection problems based on the different embedded spaces and try to select the best features from the original sets, and then improve the experimental results effectively. The main works of this paper can be described as follows:1, By introducing the ideas of the semi-supervised learning, we present a new discriminative semi-supervised nonlinear dimensionality reduction called DSSNDR for feature selection, which can be effectively calculated and can avoid the singularity. In this setting, within-class and between-class scatter information are adopted to specify whether samples belong to the same class or different classes. DSSNDR is a standard eigenvalue problem, which can be effectively computed, allows users to maximize the between-class scatters, while maintain the structure of the within-class scatters and can extract the good features based on the original sets.2, By using the domain knowledge in the form of the pairs constraints, we discuss the problems of the pairs constraints based semi-supervised nonlinear dimensionality reduction algorithms called KS2DR for feature selection, which can implement semi-supervised learning effectively. KS2DR can preserve the intrinsic structure of the unlabeled data as well as both the similar and dissimilar pairs constraints defined on the labeled training samples in the projected low-dimensional feature space, under which the samples wth different class labels are easier to be effectively partitioned from each other.3, By introducing the ideas of Linear Fisher Discriminant Analysis (FDA) and Canonical Correlation Analysis (CCA), we consider the supervised feature selection problem where samples are accompanied with class labels and propose a new Locality Preserving Multi-Vector Fisher Correlation Discriminant Analysis called LPMVF for feature selection. LPMVF takes local structure of the data into account so the multimodal data can be embedded appropriately. By defining the new guidelines, the proposed methods have some obvious advantages over FDA, KFD, and LFDA algorithms amd can effectively improve the accuracies by introducing the class labels as priori knowledge.